1 April 2021 in Lung Cancer
Deep Learning has been proposed as promising tool to classify malignant nodules. Our aim was to retrospectively validate our Lung Cancer Prediction Convolutional Neural Network (LCP-CNN), which was trained on US screening data, on an independent dataset of indeterminate nodules in an European multicentre trial, to rule out benign nodules maintaining a high lung cancer sensitivity. The LCP-CNN has been trained to generate a malignancy score for each nodule using CT data from the U.S. National Lung Screening Trial (NLST), and validated on CT scans containing 2106 nodules (205 lung cancers) detected in patients from from the Early Lung Cancer Diagnosis Using Artificial Intelligence and Big Data (LUCINDA) study, recruited from three tertiary referral centers in the UK, Germany and Netherlands. We pre-defined a benign nodule rule-out test, to identify benign nodules whilst maintaining a high sensitivity, by calculating thresholds on the malignancy score that achieve at least 99 % sensitivity on the NLST data. Overall performance per validation site was evaluated using Area-Under-the-ROC-Curve analysis (AUC). The overall AUC across the European centers was 94.5 % (95 %CI 92.6-96.1). With a high sensitivity of 99.0 %, malignancy could be ruled out in 22.1 % of the nodules, enabling 18.5 % of the patients to avoid follow-up scans. The two false-negative results both represented small typical carcinoids. The LCP-CNN, trained on participants with lung nodules from the US NLST dataset, showed excellent performance on identification of benign lung nodules in a multi-center external dataset, ruling out malignancy with high accuracy in about one fifth of the patients with 5-15 mm nodules.
28 December 2024 in European Journal of Cancer
MHMarjolein A. Heuvelmans
CACarlijn M. van der Aalst
+6
Lung cancer screening (LCS) with low-dose CT (LDCT) reduces lung-cancer-related mortality in high-risk individuals. AI can potentially reduce radiologist workload as first-read-filter by ruling-out negative cases. The feasibility of AI as first reader was evaluated in the European 4-IN-THE-LUNG-RUN (4ITLR) trial, comparing its negative-misclassifications (NMs) to those of radiologists and the impact on referral rates. NMs were collected from 3678 baseline LDCTs of the 4ITLR-dataset. LDCTs were read independently by radiologists and dedicated AI software (AVIEW-LCS, v1.1.42.92, Coreline-Soft, Seoul, Korea). A case was designated as NM when nodules > 100 mm3 were present and either radiologist or AI gave a negative-classification (only nodules 300 mm3) classification, warranting referral for clinical-workup. Overall, there were 102 referrals (2.8 %) at baseline. Of the 3678 baseline scans, 438 NMs (11.9 %) were identified (age individuals: 68 (IQR: 64-73) years, 241 men); 31 (0.8 %) by AI and 407 (11.1 %) by radiologists. Among the 31 AI-NMs, 3 were classified positive and 28 indeterminate. Among the 407 radiologist-NMs, 4 were classified positive, and 403 were indeterminate, of which 8 were classified positive after receiving a three-month follow-up CT. Radiologists, as first reader, would have led to 12/102 (11.8 %) missed referrals, higher than the 3/102 (2.9 %) of AI. This study showed AI outperforms radiologists with significantly less NMs and therefore shows promise as first reader in a LCS program at baseline, by independently ruling-out negative cases without substantially increasing the risk of missed clinical referrals.
1 April 2023 in British Journal of Radiology
Objective: To evaluate detectability and semi-automatic diameter and volume measurements of pulmonary nodules in ultralow-dose CT (ULDCT) vs regular-dose CT (RDCT). Methods: Fifty patients with chronic obstructive pulmonary disease (COPD) underwent RDCT on 64-multidetector CT (120 kV, filtered back projection), and ULDCT on third-generation dual source CT (100 kV with tin filter, advanced modeled iterative reconstruction). One radiologist evaluated the presence of nodules on both scans in random order, with discrepancies judged by two independent radiologists and consensus reading. Sensitivity of nodule detection on RDCT and ULDCT was compared to reader consensus. Systematic error in semi-automatically derived diameter and volume, and 95% limits of agreement (LoA) were evaluated. Nodule classification was compared by κ statistics. Results: ULDCT resulted in 83.1% (95% CI: 81.0–85.2) dose reduction compared to RDCT (p < 0.001). 45 nodules were present, with diameter range 4.0–25.3 mm and volume range 16.0–4483.0 mm3. Detection sensitivity was non-significant (p = 0.503) between RDCT 88.8% (95% CI: 76.0–96.3) and ULDCT 95.5% (95% CI: 84.9–99.5). No systematic bias in diameter measurements (median difference: −0.2 mm) or volumetry (median difference: −6 mm3) was found for ULDCT compared to RDCT. The 95% LoA for diameter and volume measurements were ±3.0 mm and ±33.5%, respectively. κ value for nodule classification was 0.852 for diameter measurements and 0.930 for volumetry. Conclusion: ULDCT based on Sn100 kV enables comparable detectability of solid pulmonary nodules in COPD patients, at 83% reduced radiation dose compared to RDCT, without relevant difference in nodule measurement and size classification. Advances in knowledge: Pulmonary nodule detectability and measurements in ULDCT are comparable to RDCT.